WhatsApp AI agents for restaurants and food ordering

Restaurants live on communication. Menu questions, table reservations, takeaway orders, allergy checks, delivery updates, loyalty offers — the messaging volume is relentless, and for most restaurants it's handled entirely by a member of staff who'd rather be doing something else.

WhatsApp AI agents for restaurants change that. The right setup handles routine inquiries automatically, takes orders in natural language, logs them to a system your kitchen can act on, and sends confirmation and updates without any human involvement.

TL;DR

• WhatsApp is already how customers communicate with restaurants — orders, reservations, and questions all arrive there.
• The ordering flow goes: customer messages → bot takes order in natural language → confirms total → logs to Google Sheets → sends ETA.
• CodeWords lets you build this without code — describe the agent to Cody and it handles the setup, integrations, and deployment.

Why WhatsApp for restaurants

Most restaurants already communicate with customers on WhatsApp. They just do it manually. Staff spend time copy-pasting orders, answering the same menu questions, and chasing reservation confirmations.

The volume compounds at the worst times. Friday evening — when the kitchen is at capacity — is exactly when the WhatsApp messages peak. Customers inquire about availability. Delivery customers want to know where their order is. Someone has an allergy question that needs an answer before they can place the order.

A WhatsApp AI agent handles all of this without requiring a dedicated person to sit on their phone. Staff can focus on food and service. Customers get instant, accurate responses at any hour.

Core use cases

Menu questions: a customer asks "do you do gluten-free pasta?" or "what's in the tagine?" The bot knows your full menu and can answer accurately. No waiting, no "let me check with the kitchen."

Reservations: the bot guides the customer through a booking — date, time, number of covers, any dietary requirements or special occasions. It checks availability against your calendar and confirms the booking. A reminder goes out the day before.

Takeaway ordering: customers place orders in natural conversation — "I'd like a large pepperoni pizza, a Caesar salad, and a Diet Coke." The bot understands the order, confirms it back ("That's one large pepperoni pizza, one Caesar salad, and one Diet Coke — total £24.50. Is that right?"), asks for the delivery address, and logs the order.

Allergy information: with your full allergen matrix loaded into the agent's knowledge base, it can answer any allergen question accurately. "Does the mushroom risotto contain gluten?" — "The mushroom risotto is gluten-free, but it's prepared in a kitchen where gluten is present. If you have a severe allergy, please let us know when ordering and our kitchen will take extra precautions."

Delivery updates: once an order is placed, the bot can send proactive updates — "Your order is being prepared" and "Your driver is on the way" — reducing the inbound "where's my order?" messages significantly.

Loyalty offers: the bot can send personalised WhatsApp messages to returning customers — a discount on their birthday, a special offer after their fifth visit, a preview of a new menu item for regular diners.

The ordering flow in detail

Here's how a takeaway ordering conversation looks in practice:

Customer: "Hi, I'd like to order a takeaway"

Bot: "Great — what can I get you? You can order anything from our menu. Just tell me what you'd like and I'll put it together for you."

Customer: "I'll have two chicken burgers, one plain, one with extra cheese, and large chips, and two Cokes"

Bot: "Got it — let me just confirm your order:

• 2x Chicken burger (one plain, one extra cheese) — £9.50 each
• 1x Large chips — £3.50
• 2x Coca-Cola — £2.50 each

Total: £27.50

Is that everything, or would you like to add anything?"

Customer: "That's everything"

Bot: "What's the delivery address?"

Customer: "14 Maple Street, SW4 7PQ"

Bot: "Perfect. Your order is confirmed and will be with you in approximately 35 minutes. You'll get an update when it's on its way."

At this point, the order is logged to a Google Sheet (or your POS system if integrated) and a notification goes to the kitchen. When the driver picks up, the bot sends the customer an update automatically.

The reservation flow

Reservation handling follows a similar multi-step pattern:

Customer: "Can I book a table for Saturday?"

Bot: "Of course. How many people will be dining?"

Customer: "Four"

Bot: "And do you have a preferred time? We have availability at 7pm and 8:30pm this Saturday."

Customer: "7pm works"

Bot: "Last thing — any dietary requirements or special occasions I should note?"

Customer: "One of us is vegan and it's my wife's birthday"

Bot: "Lovely — we'll make sure to note that. So that's a table for four at 7pm this Saturday, one vegan guest, celebrating a birthday. Shall I confirm that?"

Customer: "Yes please"

Bot: "Booked. We'll send you a reminder on Friday. Looking forward to seeing you."

The booking creates a record in your calendar (Calendly, Cal.com, or Google Calendar) and logs the dietary and occasion notes so the front-of-house team can prepare.

Handling allergy questions

Allergy questions are one of the most important things to get right. An inaccurate answer is a serious safety issue.

The right approach is to load your complete allergen matrix into the agent's knowledge base. This is typically a table showing which allergens (gluten, dairy, nuts, eggs, soya, celery, etc.) are present in each dish.

When a customer asks an allergen question, the bot refers to this information and gives an accurate answer. It should also always include a caveat about cross-contamination risk in shared kitchens — this is factually important and protects the business.

If a customer has a severe allergy, the bot should offer to flag this for the kitchen and suggest the customer calls to speak with a chef directly for complex dietary needs.

Example system prompt for a restaurant bot

Here's a system prompt you could describe to Cody when building your restaurant WhatsApp agent:

"You are a friendly ordering and reservations assistant for Bella Cucina, an Italian restaurant in London. Your name is Marco. You help customers with menu questions, place takeaway orders, and book tables. For takeaway orders: take the full order in natural conversation, confirm it back with prices, ask for the delivery address, log the order to our Google Sheet (columns: timestamp, customer name, phone, order items, total, address), and send a confirmation with a 35-minute estimated delivery time. For reservations: collect date, time, number of covers, and dietary requirements, check our Calendly for availability, and confirm the booking. For allergy questions: use the allergen information in the knowledge base and always mention that our kitchen handles common allergens. If a customer has a severe allergy, ask them to call us on [number] to speak with the kitchen directly."

This is exactly the kind of description you'd give to Cody (the builder inside CodeWords) in CodeWords. It's plain English — no code required.

How to build with CodeWords

CodeWords connects to Google Sheets, Calendly, Cal.com, and other restaurant-relevant tools via Composio's 3,000+ integration library. You tell Cody what you want, authenticate the integrations via one-click OAuth, and your agent goes live.

For more WhatsApp setups, visit the WhatsApp agents hub.

And if you're new to building WhatsApp agents, how to build WhatsApp AI agents with CodeWords walks through the complete process from account setup to going live.

Orders and reservations: capture rather than commit

The strong temptation with restaurants is to let the agent take the booking or the order outright. Worth resisting until the connection behind it has been built and tested properly.

The failure is specific and expensive. A confirmed table that doesn't exist, or an order that never reached the kitchen, costs you a customer and a review, and both happen when an agent confirms something it can't actually verify. The safe default is that the agent gathers everything needed, confirms what it has understood, and tells the customer when their booking or order will be confirmed by a person.

Once you have a tested connection into your booking system or point of sale, binding actions can run unattended. That's a deliberate step to take with evidence, rather than the starting position.

Allergies are their own category. An agent should never assert that a dish is free from an allergen. The correct behavior is to record the requirement, state clearly that a member of staff will confirm, and make sure that reaches the kitchen. This is one of the few places where being unhelpful is the right answer.

Where the time actually goes

Most restaurants looking at this are trying to solve one of three problems, and they need different builds.

The phone during service. Someone is answering questions about opening hours and parking while plating food. A WhatsApp agent that answers the repetitive questions takes that interruption away, and it's the easiest version to build.

Bookings arriving across four channels. Phone, WhatsApp, email, and a form, all landing in different places. The value here is consolidation into one reviewable list rather than conversation quality.

Repeat custom that never repeats. Guests who came once and were never contacted again. That's a follow-up problem, and it's governed by consent: messages to past customers need their agreement, and opt-outs need honoring permanently.

The wider back-office side of this, from supplier invoices to daily reporting, is covered in AI automation for restaurants. The WhatsApp agent is the customer-facing end of the same picture.

Frequently asked questions

Will it handle our menu changing daily?

Only if it reads the menu from wherever you maintain it rather than having it written into the agent. Specials are where this usually breaks: an agent confidently describing yesterday's dish is worse than one that says the specials change daily and offers to check.

Can it take payment?

Not something to automate early. Capture the order, have a person confirm, and keep payment on whatever channel you already trust. Anything touching payment details should be deliberate rather than convenient.

What about groups and large bookings?

Route them to a person. Large bookings involve deposits, room layout, dietary requirements, and negotiation, and they're valuable enough that the conversation is worth someone's attention.

Can it manage a waitlist on a busy night?

It can capture who's waiting and tell them where they stand, which removes a genuine source of friction at the door. Deciding who gets the next table stays with whoever is running the floor, because that call involves party size, turn times, and who has been waiting in the rain.

What happens if it gets an order wrong?

Decide the recovery before it happens, because it will. The workable pattern is that a person confirms every order before it reaches the kitchen during the first few weeks, then you narrow that review to orders above a value or complexity threshold once you have evidence it's accurate. Skipping the review period is how a small error becomes a service that nobody trusts.

Is this worth it for a single site?

Often yes, and the deciding factor is inbound volume rather than number of sites. A busy independent taking forty messages a day during service gets more from this than a small group taking five each. The question to ask is how often someone stops what they're doing to answer something they have answered a hundred times before.

How do we keep the tone right for our restaurant?

Write the tone into the instructions with examples rather than adjectives. "Friendly" means nothing to a model; two sample replies in your actual voice mean a great deal. A neighborhood bistro and a fast-casual chain should not sound the same, and the difference comes from showing rather than describing.

What should it do with complaints?

Get them to a person quickly and without argument. An agent attempting to resolve a complaint reads as a company avoiding one, and the cost of that is a public review rather than a private grumble. Acknowledge, capture the detail, and route it to whoever handles these, with the conversation attached.

Related reading

• How to build WhatsApp AI agents with CodeWords
• How to build multi-stage WhatsApp conversation flows
• How to handle voice messages in WhatsApp bots with AI